Agency at the Interface: Distinguishing Teleological from Structural Self-Organization via Internal Coarse-Graining and Downward Causation

TL;DR

The paper addresses the ambiguity in defining agency by separating observer-described models from a system's intrinsic dynamics. It introduces a duality framework where internal coarse-graining and downward causation generate self-referential intrinsic dynamics $F_S$, while an observer uses external coarse-graining and a model $F_O$, yielding a predictive/operational gap. The core contributions include formalizing the Internal and External Gaps, distinguishing teleological from structural self-organization, and outlining a spectrum of agency from minimal (e.g., bacteria) to maximal (humans) with implications for social coordination via participatory sense-making. This framework provides a mechanistic, generative basis for autonomy and suggests design principles for AI that preserve ongoing internal viability rather than mere externally guided optimization.

Abstract

Agency is widely characterized as the capacity of a system to regulate its internal states toward self-generated goals, yet characterizing the functional basis of this autonomy requires a distinction between the system's own organization and an observer's interpretation. In this paper, we ground this capacity functionally in the generative process of intrinsic dynamics, characterized as a self-referential loop generated by internal coarse-graining and downward causation. This process allows the system to autonomously compress microscopic states into macroscopic variables that subsequently constrain microscopic temporal evolution. By distinguishing the intrinsic dynamics of the system from the external coarse-graining of an observer's interpretation, we define agency through the dynamics of the predictive gap. This gap constitutes the internal divergence between the system's anticipation and its realization, as well as the limited reducibility of the system's generated constraints within the observer's interpretive model. This framework outlines a spectrum of agency that distinguishes biological systems, whose teleological constraints are generated through their own regulatory dynamics, from artificial systems, whose organization is designed to satisfy externally specified objectives. Finally, we extend this interface to the social domain, proposing that these generative divergences underpin participatory sense-making and emergent coordination.

Figures (3)

• Figure 1: A Duality Framework of Agency. The diagram shows the structural parallel between the system's intrinsic dynamics (left) and the observer's interpretive model (right). The System (blue) operates through a self-referential loop where internal coarse-graining maps microscopic states $X_t$ to intrinsic macroscopic variables $V_t^S$, which then influence the micro-level via downward causation. The Observer (black) operates through an epistemic loop where external coarse-graining maps the same microscopic states to observational variables $V_t^O$, closing the loop via interpolation. The central vertical line demarcates the interface between the system's intrinsic dynamics and the observer's epistemic description.
• Figure 2: The Duality of Gaps arising from Divergent Coarse-Graining. The diagram contrasts the temporal evolution ($t \rightarrow t'$) of the system's intrinsic dynamics (left panel, blue) with the observer's interpretive model (right panel, black). The red double-headed arrows denote the two structural divergences: the Internal Gap between the system's anticipation ($V_{t'}^{S_{ant}}$) and the realized state ($V_{t'}^{S}$), and the External Gap between the intrinsic dynamics ($F_{S}$) and the observer's model ($F{O}$). Notation: $X$ and $V$ denote microscopic and macroscopic variables, respectively. Solid arrows: Realized state transitions. Dotted arrows: Coarse-graining mappings. Dashed arrows: Anticipated or predicted trajectories.
• Figure 3: Social Interaction via the Generative Gap. (A) The reciprocal modeling loop between two agents. (B) The regime of predictive convergence. (C) The regime of emergent entanglement forms a collective macro-variable. System 1 is shown in blue; System 2 in orange. General notations for variables ($X, V$) and arrow types follow the conventions defined in Figure \ref{['fig: time evolution']}.